(AJP) A Hint of Re Normalization

A hint of renormalization
Bertrand Delamottea)
Laboratoire de Physique Theorique et Hautes Energies, Universites Paris VI, Pierre et Marie Curie, Paris VII, Denis Diderot, 2 Place Jussieu, 75251 Paris Cedex 05, France
Received 28 January 2003; accepted 15 September 2003 An elementary introduction to perturbative renormalization and renormalization group is presented. No prior knowledge of eld theory is necessary because we do not refer to a particular physical theory. We are thus able to disentangle what is specic to eld theory and what is intrinsic to renormalization. We link the general arguments and results to real phenomena encountered in particle physics and statistical mechanics. 2004 American Association of Physics Teachers. DOI: 10.1119/1.1624112
I. INTRODUCTION Hans Bethe in a seminal 1947 paper was the rst to calculate the energy gap, known as the Lamb shift, between the 2s and 2p levels of the hydrogen atom.1 These levels were found to be degenerate even in Diracs theory, which includes relativistic corrections. Several authors had suggested that the origin of the shift could be the interaction of the electron with its own radiation eld and not only with the Coulomb eld . However, to quote Bethe, This shift comes out innite in all existing theories and has therefore always been ignored. Bethes calculation was the rst to lead to a nite, accurate result. Renormalizationin its modern perturbative sensewas born.2 Since then it has developed into a general algorithm to get rid of innities that appear at each order of perturbation theory in almost all quantum eld theories QFT .37 In the meantime, the physical origin of these divergences has also been explained see Ref. 8 for many interesting contributions on the history and philosophy of renormalization and renormalization group . In QFT, as in ordinary quantum mechanics, the perturbative calculation of any physical process involves, at each order, a summation over virtual intermediate states. However, if the theory is Lorentz invariant, an innite number of supplementary states exist compared with the Galilean case and their summation, being generically divergent, produces innities. The origin of these new states is deeply rooted in quantum mechanics and special relativity. When these two theories are combined, a new length scale appears, built out of the mass m of the particles: the Compton wavelength /mc. It vanishes in both formal limits 0 and c , corresponding, respectively, to classical and Galilean theories. Because of Heisenberg inequalities, probing distances smaller than this length scale requires energies higher than mc 2 and thus imply the creation of particles. This possibility to create and annihilate particles forbids the localization of the original particle better than the Compton wavelength because the particles that have just been created are strictly identical to the original one. Quantum mechanically, these multi-particle states play a role even when the energy involved in the process under study is lower than mc 2 , because they are summed over as virtual states in perturbation theory. Thus, the divergences of perturbation theory in QFT are directly linked to its short distance structure, which is highly nontrivial because its description involves the innity of multi-particle states. Removing these divergences has been the nightmare and the delight of many physicists working in particle physics. It
170 Am. J. Phys. 72 2 , February 2004 http://aapt.org/ajp
seemed hopeless to the non-specialist to understand renormalization because it required prior knowledge of quantum mechanics, relativity, electrodynamics, etc. This state of affairs contributed to the nobility of the subject: studying the ultimate constituents of matter and being incomprehensible t well together. However, strangely at least at rst sight the theoretical breakthrough in the understanding of renormalization beyond its algorithmic aspect came from Wilsons work on continuous phase transitions.9 The phenomena that take place at these transitions are neither quantum mechanical10 nor relativistic and are non-trivial because of their cooperative behavior, that is, their properties at large distances.11 Thus neither nor c are necessary for renormalization. Something else is at work that does not require quantum mechanics, relativity, summation over virtual states, Compton wavelengths, etc., even if in the context of particle physics they are the ingredients that make renormalization necessary. In fact, even divergences that seemed to be the major problem of QFT are now considered only as byproducts of the way we have interpreted quantum eld theories. We know now that the invisible hand that creates divergences in some theories is actually the existence in these theories of a no mans land in the energy or length scales for which cooperative phenomena can take place, more precisely, for which uctuations can add up coherently.12 In some cases, they can destabilize the physical picture we were relying on and this manifests itself as divergences. Renormalization, and even more renormalization group, is the right way to deal with these uctuations. One of the aims of this article is to disentangle what is specic to eld theory and what is intrinsic to the renormalization process. Therefore, we shall not look for a physical model that shows divergences,1317 but we shall rather show the general mechanism of perturbative renormalization and the renormalization group without specifying a physical model. II. A TOY MODEL FOR RENORMALIZATION In the following, we consider an unspecied theory that involves, by hypothesis, only one free parameter g 0 in terms of which a function F(x), representing a physical quantity, is calculated perturbatively, that is, as a power series. An example in QFT would be quantum electrodynamics QED , which describes the interaction of charged particles such as electrons with the electromagnetic eld. For high energy processes, the mass of the electron is negligible and the only parameter of this theory in this energy regime is its charge,
2004 American Association of Physics Teachers 170
which is therefore the analog of g 0 . F can then represent the cross section of a scattering process as, for instance, the scattering of an electron on a heavy nucleus in which case x is the energymomentum four-vector of the electron. The coupling constant g 0 is dened by the Hamiltonian of the sys` tem, and F is calculated perturbatively using the usual a la Feynman approach. Another important example is continuous phase transitions. For uids, F could represent a densitydensity correlation function and for magnetism a spinspin correlation function.18 Yet another example is the solution of a differential equation that can arise in some physical context and that can show divergences see the following . It is convenient for what follows to assume that F(x) has the form: F x g0 g 2F 1 0 x g 3F 2 0 x . 1
dened expansion. The hypothesis is therefore that the problem does not come from the perturbation expansion itself, that is, from the functions F i (x), but from the choice of parameter used to perform it. This hypothesis means that the physical quantity, F(x), initially represented by its illdened expansion Eq. 1 , should have a well-dened perturbation expansion once it is calculated in terms of the physical parameter F( ). This is the simplest hypothesis we can make, because it amounts to preserving the x-dependence of the functions F i (x) and only modifying the coupling constant g 0 . Thus, we assume that F(x) is known at one point , and we dene g R by F gR . 3
Up to a redenition of F, this form is general and corresponds to what is really encountered in eld theory. Let us now assume that the perturbation expansion of F(x) is illdened and that the F i (x) are functions involving divergent quantities. An example of such a function is F1 x dt , t x 2
which is logarithmically divergent at the upper limit. This example has been chosen because it shares many common features with divergent integrals encountered in QFT: the integral corresponds to the summation over virtual states and (t x) 1 represents the probability amplitude associated with each of these states.19 A simple although crucial observation is that because there is only one free parameter in the theory by hypothesis, only , is necone measurement of F(x), say at the point x essary to fully specify the theory we are studying. Such a measurement is used to x the value of g 0 so as to reproduce the experimental value of F( ). For QED for instance, this procedure would mean that: i ii iii We start by writing a general Hamiltonian compatible with basic assumptions, for example, relativity, causality, locality, and gauge invariance. We calculate physical processes at a given order of perturbation theory. We x the free parameter s of the initial Hamiltonian to reproduce at this order the experimental data.
In the following, and by analogy with QFT, we call g R the renormalized coupling constant and Eq. 3 a renormalization prescription, a barbarian name for such a trivial operation. We are now in a position to discuss the renormalization program. It consists of reparametrizing the perturbation expansion of F so that it obeys the prescription of Eq. 3 . The point here is that we cannot use Eq. 3 together with Eq. 1 because Eq. 1 is ill-dened. We rst need to give a welldened meaning to the perturbation expansion. This is the regularization procedure which is the rst step of any renormalization.20,21 The idea is to dene the perturbation expansion of F by a limit such that i the F i (x) are welldened before the limit is taken, and ii after the renormalization has been performed, the original formal expansion is recovered when the limit is taken. We thus introduce a new set of regularized functions F and F i, , involving a new parameter , which we call the regulator, and such that for nite all these functions are nite. We thus dene F x F x,g 0 , g 0 g 2 F 1, x 0 g 3 F 2, x 0 . 4
There are innitely many ways of regularizing the F i s and for the example given in Eq. 2 , it can consist for instance in introducing a cut-off in the following integral: F 1, x dt . t x 5
This last step requires as much data as there are free parameters. Once the parameters are xed, the theory is completely determined and thus predictive. One could then think that it does not matter whether we parametrize the theory in terms of g 0 , which is only useful in intermediate calculations, or with a physical, that is, a measured quantity F( ), because g 0 will be replaced by this quantity anyway. Having this freedom is indeed the generic situation in physics, but the subtlety here is that the perturbation expansion of F(x) is singular, and, thus, so is the relationship between g 0 and F( ). Thus, it seems crucial to reparametrize F in terms of F( ) when the expansion is ill-dened. The renormalizability hypothesis is that the reparametrization of the theory in terms of a physical quantity, instead of g 0 , is enough to turn the perturbation expansion into a well171 Am. J. Phys., Vol. 72, No. 2, February 2004
Different regularization schemes can lead to very different intermediate calculations, but must all lead to identical results.22 For instance, dimensional regularization is widely used in QFT because it preserves Lorentz and gauge symmetries.1315,23 We do not need here to specify a regularization for the function F, because our arguments will be general and the few calculations elementary. Once a regularization scheme has been chosen, it is possible to use the renormalization prescription, Eq. 3 , together with the regularized expansion, Eq. 4 , to obtain a well-dened perturbation series for F in terms of the physical coupling g R . If this expansion makes sensethis is the renormalizability hypothesisit must be nite even in the limit , because it expresses a nite physical quantity F(x) in terms of a physical quantity g R . Thus, the renormalization program consists rst in changing F(x,g 0 ) to F (x,g 0 , ), then in rewriting F in terms of g R and , F x,g 0 , F x,g R , ,
Bertrand Delamotte
6
171
and only then taking the limit at xed g R and . If this limit exists, F (x) is by hypothesis the function F(x): F x F x,g R , F
F x
gR
2 x gR
dt t x t
3 O g R , 16
x,g R ,
Of course, the divergences must still be somewhere, and we shall see that they survive in the relationship between g 0 and g R ; at xed g R , g 0 diverges when . In the traditional interpretation of renormalization, this divergence is supposed to be harmless because g 0 is supposed to be a nonphysical quantity. We shall come back to this point later. The renormalization program is performed recursively, and we now implement it order by order to see how it works and the constraints on the perturbation expansion that it implies. Let us emphasize that the series expansion we shall use in intermediate calculations are highly formal because they are ill-dened in the limit . They are justied only by the result we nally obtain: a good perturbation expansion in terms of g R . 24 Renormalization at order g 0 . At this order F(x) is constant and given by F x g0 O g2 . 0 8
Thus the use of Eq. 3 leads to

2 g0 gR O gR .
9 g2 . 0
Renormalization at order Our only freedom to eliminate the divergence of F (x) is to redene g 0 . Because we are working perturbatively, we expand g 0 as a power series in g R . Thus, we set g0 gR where F
ng 2g n O(g R ). 3g
,
2 gR
10 we obtain
3 O gR ,
At order
2 g R F 1,
gR
2g
11
where we have used at this order, we obtain

2g 2 g R F 1,
g2 0 ,
2 gR
3 O(g R ).
If we impose Eq. 3 12
which diverges when nd

2g 2 gR
. In our example, Eq. 5 , we dt

2 g R log
which is obviously well dened and such that the prescription of Eq. 3 is veried. We say that we have renormalized the theory to this order. Before going to the next order of perturbation theory, let us note two important facts. First, the renormalization procedure consists of adding a divergent term 2 g to F to remove its divergence. The cancellation takes place between the second term of its expansion and the rst one of order 2 g 0 . Both lead to a term of order g R , the one coming from the expansion of g 0 in terms of g R being tuned so as to cancel the divergence of the other. This mechanism of cancellation is a general phenomenon: a divergence coming from the nth term of the perturbation expansion is canceled by the expansion in powers of g R of the n 1 preceding terms. Second, this cancellation is possible for all x only if the divergence of F 1, (x) is a number, that is, is x-independent. If it is not so, then F 1, (x) F 1, ( ) would still be divergent x . This divergence would require the imposition of at least one more renormalization prescription to be removed and this second prescription would dene a second, independent, coupling constant see Appendix A for two functions, one renormalizable and one that is not . The necessity for a second measurement of F(x) would contradict our assumption that there is only one free parameter in the theory. Thus we conclude that this assumption drastically constrains the x-dependence of the divergences at order g 2 . 0 We actually show in the following that this constraint propagates to any order of perturbation theory in a nontrivial way. We also will show that together with dimensional analysis and for a very wide and important class of theories, these constraints are sufcient to determine the analytical form of the divergences. Renormalization at order g 3 . We suppose that F can be 0 2 renormalized at order g R , that is, condition 15 is fullled. To understand the structure of the renormalization procedure, 3 it is necessary to go one step further. At order g R we obtain F x gR
2g 3g 2 g R 2g R 2g
F 1, x 17
13
3 g R F 2,
x g3 0
O
3 gR
4 gR
, g2 0
2 gR
If we substitute Eq. 12 into Eq. 11 , we obtain F to this order: F x

2 g R g R F 1, x
and 2g R 2 g where we have used 4 O(g R ). We again impose the prescription Eq. 3 and obtain
3g 3 2g R F 1, 2 3 g R F 2,
4 O(g R )
F 1,
3 O gR .
14
18
It is clear that this expression for F (x) is nite for all x at this order if and only if the divergent part of F 1, (x) the part that becomes divergent when ) is exactly canceled by that of F 1, ( ), that is, if and only if F 1, x F 1, is regular in x and for . 15 This condition of course means that the divergent part of F 1, (x) must be a constant, that is, is x-independent. If this is so, then we dene the function F(x)now called renormalizedas the limit of F (x) when . The condition 15 is fullled for the example of Eq. 2 , and we trivially nd that F(x) reads:
172 Am. J. Phys., Vol. 72, No. 2, February 2004
If we substitute Eq. 18 in Eq. 17 , we obtain F x

2 g R g R F 1, x
F 1, F 1, x
3 g R F 2, x
F 2, O
4 gR
2F 1, .
F 1, 19
Once again, we require that the divergence has been subtracted for all x which imposes on the x-dependence of the divergent part of F 2, (x): F 2, x F 1, F 2, 2F 1, F 1, x as . 20
172
is regular in x and
Bertrand Delamotte
Note that this constraint involves not only F 2, , but also F 1, . It is convenient to rewrite F 1, (x) and F 2, (x) as the sum of a regular and of singular when ) part: F i, x
s F i,
x-independent. In our example of Eqs. 2 and 24 we nd

s F 2, x
log
log
x x
f2
28
r F i,
x .
21
Because anything nite , this decomposition is not s unique: the F i, (x) are dened up to a regular part. It is s convenient to choose F 1, (x) such that
s F 1,
1 and by By expanding log log(x )/ x in powers of again redening the regular part of F 2, , we obtain a simpler s form for F 2, (x): s F 2, x
log
log x
f2
29
s F 1,
0,
22
which, of course, implies condition 15 . We show in Appendix B that, reciprocally, this choice is always possible if 15 is fullled. As already stated, Eq. 22 means that the divergent part of F 1, is x-independent. We can actually impose a s more stringent condition on F 1, because, by again tuning s the regular part of F 1, , we can choose F 1, to be completely independent of x, for any . We thus dene
s F 1, x
f1
23
In our example, Eq. 5 , we can choose f1 log ,

r F 1, x
log
x x
24
We now substitute Eq. 23 into Eq. 20 and, using the same kind of arguments as in Appendix B, we obtain a constraint s on the singular part of F 2, (x) similar to the one on F 1, (x), Eq. 22 :
s F 2, x s F 2,
2 f1
r F 1, x
r F 1,
0.
25 Equation 25 can be rewritten as

s F 2, x s F 2, r F 1, x
2 f1
2 f1
r F 1,
0.
26
Equation 26 has the same structure as Eq. 22 up to the s s r replacement: F 1, F 2, 2 f 1 ( )F 1, and therefore has the same kind of solution as Eq. 23 :
s F 2, x
2 f1
r F 1, x
f2
27
where f 2 ( ) is any function of and is independent of x. We see in Eq. 27 that unlike F 1, (x), the divergent part of F 2, (x) depends on x. However, this dependence is entirely determined by the rst order of the perturbation expansion. The 2 g term, necessary to remove the O(g 2 ) divergence, 0 3 has produced at order g R an x-dependent divergent term: 2g R 2 gF 1, (x). This kind of x-dependence is also a general phenomenon of renormalization: the counter- terms that remove divergences at a given order produce divergences at higher orders. If the theory is renormalizable, these divergences contribute to the cancellation of divergences present in the perturbation expansion at higher orders. Thus, perturbative renormalizability, that is, the possibility of eliminating order by order all divergences by the redenition of the coupling s , implies a precise structure of the divergent parts of the successive terms of the perturbation series. At order n, the singular part of F n, involves x-dependent terms entirely determined by the preceding orders plus one new term that is
This relation will be important in the following when we shall discuss the renormalization group. Let us draw our rst conclusion. Innities occur in the perturbation expansion of the theory because we have assumed that it was not regularized. Actually, these divergences have forced us to regularize the expansion and thus to introduce a new scale . Once regularization has been performed, renormalization can be achieved by eliminating g 0 . The limit can then be taken. The process is recursive and can be performed only if the divergences possess, order by order, a very precise structure. This structure ultimately expresses that there is only one coupling constant to be renormalized. This means that imposing only one prescripis enough to subtract the divergences for all x. tion at x In general, a theory is said to be renormalizable if all divergences can be recursively subtracted by imposing as many prescriptions as there are independent parameters in the theory. In QFT, these are masses, coupling constants, and the normalization of the elds. An important and non-trivial topic is thus to know which parameters are independent, because symmetries of the theory like gauge symmetries can relate different parameters and Green functions . Let us once again recall that renormalization is nothing but a reparametrization in terms of the physical quantity g R . 25 The price to pay for renormalizing F is that g 0 becomes innite in the limit , see Eq. 12 . We again emphasize that if g 0 is believed to be no more than a non-measurable parameter, useful only in intermediate calculations, it is indeed of no consequence that this quantity is innite in the limit . That g 0 was a divergent non-physical quantity has been common belief for decades in QFT. The physical results given by the renormalized quantities were thought to be calculable only in terms of unphysical quantities like g 0 called bare quantities that the renormalization algorithm could only eliminate afterward. It was as if we had to make two mistakes that compensated for each other: rst introduce bare quantities in terms of which everything was innite, and then eliminate them by adding other divergent quantities. Undoubtly, the procedure worked, but, to say the least, the interpretation seemed rather obscure. Before studying the renormalization group, let us now specialize to a particular class of renormalizable theories. III. RENORMALIZABLE THEORIES WITH DIMENSIONLESS COUPLINGS A very important class of eld theories corresponds to the situation where g 0 is dimensionless, and x, which in QFT represents coordinates or momenta, has dimensions or more generally when g 0 and x have independent dimensions . In four-dimensional spacetime, quantum electrodynamics is in this class, because the ne structure constant is dimensionless; quantum chromodynamics and the WeinbergSalam
Bertrand Delamotte 173
model of electro-weak interactions are also in this class. In four space dimensions, the 4 model relevant for the GinzburgLandauWilson approach to critical phenomena is in this class too. This particular class of renormalizable theories is the cornerstone of renormalization in eld theories. Our main goal in this section is to show that, independently of the underlying physical model, dimensional analysis together with the renormalizability constraint determine almost entirely the structure of the divergences. This underlying simplicity of the nature of the divergences explains that there is no combinatorial miracle of Feynman diagrams in QFT as it might seem at rst glance. Let us now see in detail how it works. Because F (x) has the same dimension as g 0 , it also is dimensionless and so are the F i, (x). The only possibility for a dimensionless quantity like F to be a function of a dimensional variable like x is that there exists another dimensional variable such that F depends on x only through the ratio of these two variables. Apart from x, the only other quantity on which F depends is , which must therefore have the same dimension as x. This is indeed the case in our example, Eq. 5 . Thus, the functions F i, (x) depend on the ratio x/ only.26 Let us show that this is enough to prove s that the F i, (x) are sums of powers of logarithms with, for most of them, prescribed prefactors. s Let us start with F 1, (x). On one hand, we have seen that by redening the regular part of F 1, (x), we could take its s singular part F 1, (x) independent of x, Eq. 23 . On the other hand, we know that F 1, (x) is a function of x/ . Thus, r by redening F 1, (x), it must be possible to extract an x-dependent regular part, r(x), of this function so as to build s the x/ dependence of F 1, (x):
s F 1, x s F 1,
dimensional analysis. We have already partially studied this s case with the example given in Eq. 5 where F 1, (x) is logarithmically divergent, a characteristic feature of these renormalizable theories. In particular, we have shown that in s this case, renormalizability imposes at order g 3 that F 2, is 0 of the form given in Eq. 29 . Let us now use dimensional s analysis that once again imposes that F 2, depends only on x/ . The only freedom we have to reconstruct a function of s x/ from the form of F 2, given in Eq. 29 is to add a regular function to it. It is not difcult to nd how to proceed because the only admissible term including log log x is log2 /x:
log2
log2
2 log
log x log2 x.
32
Thus, to obtain the dimensionally correct extension of the term 2 2 log log x in Eq. 29 , we extract 2 log2 from f 2 ( ) and add the regular term 2 log2 x: 2
2
log
log x
2 2
f2 log x log x
2 2
2 2
2
log log
log2 log2
f2
2
log2 f2
log2 x
log2 log2 x f2
2
log2
33
s Thus, we obtain for the new function F 2, (x):
f1
r x .
30
s F 2, x
log2
f2
log2
34
is separable into functions of x only and of Hence, only which sum up to a function of x/ . We show in Appendix C the well-known fact that only the logarithm obeys this property. We obtain see Eqs. C3 and C4
s F 1, x
f1
f1
f1 x
log
31
Therefore, for renormalizable theories and for dimensionless functions such as F, only logarithmic divergences are allowed at order g 2 in QFT, this is the so-called one-loop 0 term . This is the reason why logarithms are encountered everywhere in QFT. Note that because of dimensional analysis, the nite part of F 1, (x) is nothing but r(x), up to an additive constant, at least for . This can be checked for the example given in Eq. 5 . Thus, by dimensional analysis, the structure of the divergence determines that of the nite part up to a constant . Notice that things would not be that simple if F (x) depended on another dimensional parameter, which is the case of massive eld theories where masses and momenta have the same dimension. In this case, the nite part is only partially determined by the singular one. s Let us now show that the structure of F 2, also is entirely determined for renormalizable theories with dimensionless couplings both by the renormalizability hypothesis and by
2 log2 , we can repeat the same arguNow, for f 2 ( ) s ment as the one used previously for F 1, (x) which is equal to f 1 ( ), Eq. 23 : it is a function of that must become a function of x/ only by adding a function of x. It is thus also a logarithm, see Eqs. 30 and 31 and Appendix C. s Therefore, we add a log x term to F 2, (x) and obtain the nal result:
s F 2, x
log2
log
35
where is a pure number. We emphasize that although it is s x-independent, the term 2 log2 involved in F 2, (x) arises from the log log x term thanks to dimensional analysis. It is thus entirely determined by the term of order g 2 of perturba0 tion theory. Only the sub-leading logarithm log /x is new. It is not difcult now to guess the structure of the next order of perturbation: it involves a log3 /x with a prefactor 3 , a log2 /x term with a prefactor which is a function of and and a log /x with a prefactor independent of and . A precise calculation shows that
Fs x
g 2 log 0
2 3 g0
log2
x 5 2
3 4 g0
log3
x x
36
By substituting this expression in Eqs. 36 and 39 , we 4 obtain at O(g R ): F x gR

2 g R log
g 3 log 0
g 4 log2 0
2 3 gR
log2
x 5 2
3 4 gR
log3
g 4 log 0
3 g R log
4 g R log2
We have written the series so as to exhibit its triangular nature: the rst line corresponds to the leading logarithms, the second to the sub-leading, etc., and the nth column to the (n 1)th order of perturbation. The leading logarithms are entirely controlled by the g 2 term, the sub-leading logarithms 0 by both the g 2 and g 3 terms, etc. It is clear that order by 0 0 order for the divergent terms, only the log term is new, all the log2, log3, etc., terms are determined by the preceding orders. This structure strongly suggests that we can, at least partially, resum the perturbation series. We notice that although the leading logarithms form a simple geometric series, this is no longer true for the sub-leading logarithms where, for instance, the factor 5 /2 of Eq. 36 is non-trivial. Thanks to the renormalization group, there exists a systematic way to perform these resummations27 see the following . We again emphasize that for our simple toy model the divergences together with dimensional analysis determine almost entirely the entire function F(x) in the limit of large . To show this explicitly, we rewrite F as F x,g 0 ,
s
4 g R log
. 41
Thus, we nd that the renormalization process leaves unchanged the functional form of F , Eq. 36 , and just consists in replacing (g 0 , ) by (g R , ). This very important fact is related to a self-similarity property that we study in detail from the renormalization group viewpoint. Notice that of course any explicit dependence on and g 0 has been eliminated in Eq. 41 and that the limit can now be safely taken, if desired. Note that we have obtained logarithmic divergences because we have studied the renormalization of a dimensionless coupling constant. If g 0 was dimensional, we would have obtained power law divergences. This is for instance what happens in QFT for the mass terms see also in the following the expansion in Eq. 45 . IV. RENORMALIZATION GROUP Although the renormalization group will allow us to partially resum the perturbation expansion, we shall not introduce it in this way. Rather, we want to examine the internal consistency of the renormalization procedure. We have chosen a renormalization prescription at the point where g R is dened. Obviously, this point is not spex cial, and we could have chosen any other point or to parametrize the theory. Because there is only one independent coupling constant, the different coupling constants g R g R ( ), g R g R ( ), g R g R ( ) should all be related in such a way that F(x) F(x, ,g R ) F(x, ,g R ) F(x, ,g R ), etc. This means that there should exist an equivalence class of parametrizations of the same theory and that it should not matter in practice which element in the class is chosen. This independence of the physical quantity with respect to the choice of prescription point also means that the changes of parametrizations should be a renormalization group law: going from the parametrization given by ( ,g R ) to that given by ( ,g R ) and then to that given by ( ,g R ) or going directly from the rst parametrization ( ,g R ) to the last one ( ,g R ) should make no difference, see Fig. 1. Put this way, this statement seems to be void. Actually, it is. More precisely, it would be so if we were performing exact calculations: we would gain no new physical information by implementing the renormalization group law. This is because this group law does not reect a symmetry of the physics, but only of the parametrization of our solution. This situation is completely analogous to what happens for the solution of a differential equation: we can parametrize it at time t in terms of the initial conditions at time t 0 for instance, or we can use the equation itself to calculate the solution at an intermediate time and then use this solution
g0 F
x,g 0 ,
x,g 0 , O(g 4 ) 0
37
with F (x,g 0 , ) given by Eq. 36 at and F r (x,g 0 , ) O(g 2 ). From dimensional analysis, 0 F r (x,g 0 , ) is also a function of x/ only which, by denition, is nite when . Thus, for large , F r x,g 0 , F x ,g 0 F 0,g 0 . 38
F r (x) is therefore almost x-independent for large : it is a (g 0 -dependent number in this limit. For the sake of simplicity, let us consider the case where it is vanishing: F x,g 0 , g 0 F s x,g 0 , 39 with F s (x,g 0 , ) a function of x/ only. By using the renormalization prescription, Eq. 3 , we can calculate g R as a function of g 0 and / and by formally inverting the se4 ries, we obtain at O(g R ): g0 gR
4 gR 2 g R log 3 gr 2
log2 log2
log
3
log
5 2
log3
40
Fig. 1. An illustration of the renormalization group: the two equivalent ways to compose changes of parametrizations. 175 Am. J. Phys., Vol. 72, No. 2, February 2004
as a new initial condition to parametrize the solution at time t. The changes of initial conditions that preserve the nal solution can be composed thanks to a group law. Let us consider, for example, the following trivial, but illuminating, example: y t y t y t , y t0 r0 , r 0e
(t t 0 )
y t
r0 1
t t0
t t0
51
42 . , 43 44
the solution of which is f r 0 ,t t 0 The group law can be written as f r 0 ,t t 0 f f r0 ,

28
t 0 ,t
which you can verify using the exact solution, Eq. 43 . The non-trivial point with these group laws is that, in general, they are violated at any nite order of the perturbation expansions. In our previous example, we obtain to order , y t and f 1 f 1 r0 , t 0 ,t r0 1
2
f 1 r 0 ,t t 0
r0 1
t t0 , t t0 r0 t t0 .
45
46
The group law is veried to order because the perturbation expansion is exact at this order. However, it is violated by a term of order 2 that can be arbitrarily large even for small , provided t t 0 is large enough. The interest of the group law, Eq. 44 , is that it is possible to enforce it and then to improve the perturbation result. Actually, when renormalization is necessary, the group law lets us partially resum the perturbation series of divergent terms. Let us now see how this improvement of the perturbation series works for the example of the differential equation 42 . . Thus, t 0 In this case, the divergence occurs for t 0 plays the role of the cut-off , t t 0 of log / , and t of log / . Once t 0 is nite, no divergence remains, but the relics of the divergences occurring for t 0 are the large violations of the group law because both the divergences and these violations originate in the fact that the perturbation expansion is performed in powers of (t t 0 ) and not of . To further study the relevance of the group law, it is interesting to forget the higher order terms of the perturbation expansion for a while and to look for an improved approximation that coincides at order with the perturbation result and that obeys the group law at order 2 : f imp r 0 ,t t 0 1 r0 1 t t0
2
Thus, the rst order in the perturbation expansion, together with the group law, determines entirely the term of highest degree in t t 0 at the next order. Of course, to verify exactly the group law, we should pursue the expansion in to all orders. It is easy to show that to order n , the term of highest degree in t t 0 is completely determined by both the rstorder result and the group law and coincides with the perturbation result: n (t t 0 ) n /n!. Thus, the only information given by the perturbation expansion is that all subdominant terms, n (t t 0 ) p with p n, vanish in this example. We could now show how the implementation of the group law lets us resum the perturbation expansion. Unfortunately, this example is too simple and some important features of the renormalization group are missed in this case. See Appendix E for a complete discussion of the implementation of the renormalization group on this example. We therefore go back to our toy model for which we specialize to renormalizable theories with dimensionless couplings. Renormalization group for renormalizable theories with dimensionless couplings. We now reconsider our toy model, Eqs. 4 , 36 , and 37 , from the point of view of the renormalization group. For the sake of simplicity, we keep only the dominant terms at each order, that is, apart from g 0 , the divergent ones in Eq. 39 . First, notice that in the same way g R is clearly associated with the scale , Eq. 3 , so is g 0 with the scale because from Eq. 36 , we nd29 F x g0 . 52 Let us dene a third coupling constant associated with the scale , F gR , 53
and study the relationship between these different coupling constants at order g 2 . From 0 F x,g 0 , g0 g 2 log 0 x O g3 , 0 54
we obtain gR g0 gR g0 g 2 log 0 g 2 log 0 O g3 , 0 O g3 . 0 55 56
G t t0 .
2
47
By imposing the group law, Eq. 44 , to order a functional equation for G: G t t0 G t0 G t t0 t
, we obtain . 48
By eliminating g 0 between these two equations, we nd gR gR

2 g R log 3 O gR ,
57
If we differentiate Eq. 48 with respect to t 0 and take t 0 , we obtain, setting x t , G x x G 0 . x2 2 49
Because G(0) 0, Eq. 49 implies that G x ax, 50
where a is arbitrary. For a 0, this result is actually the perturbation result to order 2 because
and thus, as expected, the group law controlling the change of prescription point is veried perturbatively. We note that the essential ingredient for this composition law is that Eq. 57 is independent of . This is what ensures that the same form can be used to change (g 0 , ) into (g R , ) and then (g R , ) into (g R , ). This independence, in turn, is nothing but the signature of perturbative renormalizability which lets us completely eliminate at each order (g 0 , ) for (g R , ). Perturbatively, everything looks ne. However, the previous calculation relies on a formal step that is not mathematically
correct, at least for large . Indeed, to go from Eq. 56 to Eq. 57 , the series g R g R (g 0 ) must be inverted to nd g 0 g 0 (g R ) while, for , the series g R g R (g 0 ) is clearly not convergent and thus not invertible. Thus, the neglected 3 terms of order g R in Eq. 57 involve a term proportional to )( t 0 ) of log / log / analogous to the term (t Eqs. 46 and 48 which is neglected because it is of order 3 see Appendix D . g R , but which is very large for large From a practical point of view, the existence at any order of these large terms of higher orders spoil the group law so that the independence of the physical results with respect to the choice of prescription point is not veried. As in the case of the differential equation 47 , we can look for an improved function: F imp, F imp x,g 0 , g0 g 2 log 0 g 3G 0 O g4 , 0 58 for which the group law at order g 3 is obeyed. It is shown in 0 Appendix D that this constraint implies that G x where
2
f gR ,
f f gR ,
62
log2 x
log x,
59
is arbitrary. Thus, g0 g 2 log 0 g 3 log 0

2 3 g0
F imp x,g 0 ,
log2
x 60
O g4 . 0
Once again, we nd that the group law together with the order g 2 result determines the leading behavior at the next 0 order, here the log2( /x) term. Moreover, we nd that the group law imposes the existence of the same log2 term as the one found from the renormalizability constraint, Eqs. 35 and 36 , and allows the existence of a sub-leading logarithm. Although nontrivial, this should not be too surprising because the renormalizability constraint means that once F is , it also is everywhere and in particular well dened at x at x . The renormalizability constraint is therefore certainly necessary for the implementation of the group law. As in the example of the differential equation, Eq. 42 , we should pursue the expansion to all orders to obtain an exactly veried group law. It is clear that by doing so, we would nd the same expansion as the one obtained from the renormalizability constraint. Thus, if we use perturbation theory to calculate the coefcient in front of the rst leading logarithm of order g 2 ) and impose the group law, we should be able to 0 resum all the leading logarithms. To do the resummation of the sub-leading and sub-sub-leading logarithms, a knowledge of, respectively, the order g 3 and g 4 terms is required. 0 0 Clearly, we need to understand how to systematically construct the function f giving g R in terms of g R and / , 27 gR f gR , , 61
The function f is then said to be the self-similar approximation at order n of the exact relationship between g R and g R . 30 First notice one crucial thing. Our rst aim was to study the perturbation expansion of a function F in a power series of a coupling constant g 0 . Then we have discovered that the logarithmic divergence at order g 2 propagates to all 0 orders so that the expansion is actually performed in is the regulator, it is g 0 log / instead of g 0 . Because supposed to be very large compared with , so that the large logarithmic terms invalidate the use of the perturbation expansion. Reciprocally, it is clear that perturbation theory is perfectly valid if it is performed between two scales 1 and 2 which are very close. Thus, instead of using perturbation theory to make a big jump between two very distinct scales, say and , we should use it to perform a series of very little steps for which it is valid at each of them. In geometrical terms, the fact that the perturbative approach is valid only between two very close scales means that we should not use perturbation theory to approximate the equation of the curve given by the function f , Eq. 61 , that joins the points ( ,g R ) and ( ,g R ), but we should use it to calculate the eld of tangent vectors to this curve, that is, its envelope. The curve itself should then be reconstructed by integration, see Appendix E. By doing so, the group law will be automatically veried because, by construction, the integration precisely consists in composing innitesimal changes of reparametrization innitely many times. Let us consider again Eq. 55 . We want to calculate the evolution of g R ( ) with for a given model specied by ( ,g 0 ). Thus we dene gR gR
g0 ,
63
which gives the innitesimal evolution of the coupling constants with the scale for the model corresponding to (g 0 , ). We trivially nd to this order from Eq. 55 , gR g2 O g3 , 0 0 64 and thus, by trivially inverting the series of Eq. 55 , we obtain gR
2 3 gR O gR .
65
Now, if we integrate Eq. 63 together with Eq. 65 , we obtain gR 1 gR g R log . 66
such that its expansion at order n is given by the nth order of perturbation theory, the group law is exactly veried:
This relation has several interesting properties. 2 i When expanded to order g R , the perturbation result to this order is recovered, Eq. 57 . This is quite normal because (g R ) has been calculated to this order. ii When expanded to all orders, the whole series of leading logarithms is recovered. This is more interesting because (g R ) has been calculated only to order g 2 , but simply means that all the leading logarithms are determined by the rst one. iii The group law 62 is obeyed exactly. We have thus found the function f of Eq. 61 to this order. It is very
instructive to check the group law directly from Eq. 66 and to verify that the function found in Eq. 65 is not modied if we add the leading logarithmic term of order g 3 to relation 0 55 : gR g0 g 2 log 0
2 3 g0
1 gR
1 gR
log
gR gR
gR gR
log
71
log2
O g4 . 0
67
The independence of the function with respect to the addition of the successive leading logarithmic terms means that this function is indeed the right object to build selfsimilar approximations out of the perturbation expansion. Let us now return to the function itself. First, we have calculated the logarithmic derivative gR / instead of the ordinary derivative with respect to because we wanted to have a dimensionless function. Second, even the dimensionless quantity, (g R ), could have depended on / . However, the evolution of g R ( ) between and d cannot depend in perturbation theory on because the theory is perturbatively renormalizable: the perturbative relation between g R ( ) and g R ( ) depends only on and and not on . Thus, being dimensionless, the function cannot depend on alone and is thus only a function of g R . This property is general for any renormalizable theory: in the space of coupling constants, the function is always a local function. Third, the function is the function to be expanded in perturbation theory because it is given by a true series in g R and not in g R log / . This is clear for our example, Eq. 65 , where there is no logarithm, and can be proven formally by the following argument. If we use Eqs. 61 and 63 , we nd that gR f g ,y y R
y 1
There is no simple solution of this transcendental equation. It is however possible to obtain an iterative solution that is 3 2 valid if the O(g R ) term is small compared with the O(g R ) one, that is, if g R / 1. It is obtained by replacing g R in the third term of Eq. 71 by its expression obtained to order 2 g R , Eq. 66 : gR 1 g R log gR g R log 1 g R log . 72
It is easy to check that Eq. 72 resums exactly all the leading and sub-leading logarithms of the perturbation expansion Eq. 41 . Note that contrary to the one-loop result, Eq. 66 , which resums only the leading logarithms, the exact expression in Eq. 71 contributes also to the sub-sub-leading logarithms as well as the sub-sub-sub-leading ones and so on and so forth. V. SUMMARY 1 The long way of renormalization starts with a theory depending on only one parameter g 0 , which is the small parameter in which perturbation series are expanded. In particle physics, this parameter is in general a coupling constant like an electric charge involved in a Hamiltonian more precisely the ne structure constant for electrodynamics . This parameter is also the rst order contribution of a physical quantity F. In particle/statistical physics, F is a Green/ correlation function. The rst order of perturbation theory neglects uctuationsquantum or statisticaland thus corresponds to the classical/mean eld approximation. The parameter g 0 also is to this order a measurable quantity because it is given by a Green function. Thus, it is natural to interpret it as the unique and physical coupling constant of the problem. If, as we suppose in the following, g 0 is dimensionless, so is F. Moreover, if x is dimensionalit represents momenta in QFTit is natural that F does not depend on it as is found in the classical theory, that is, at rst order of the perturbation expansion. 2 If F does depend on x, as we suppose it does at second order of perturbation theory, it must depend on another dimensional parameter, , through the ratio of x and . If we have not included this parameter from the beginning in the model, the x-dependent terms are either vanishing, which is what happens at rst order, or innite as they are at second and higher orders. This is the very origin of divergences from the technical point of view . 3 These divergences require that we regularize F. This requirement, in turn, requires the introduction of the scale that was missing. In the context of eld theory, the divergences occur in Feynman diagrams for high momenta, that is, at short distances. The cut-off suppresses the uctua1 . In statistical tions at short distances compared with physics, this scale, although introduced for formal reasons, has a natural interpretation because the theories are always effective theories built at a given microscopic scale. It corresponds in general to the range of interaction of the constituents of the model, for example, a lattice spacing for spins, the average intermolecular distance for uids. In particle
. ),
68
If f is a double series in g and in log( / f gR ,

n n,p g R
logp
69
n,p
it is clear from Eq. 68 that only terms with p 1 contribute to (g), with the logarithm replaced by 1. Thus we immediately deduce from this argument and from Eq. 36 that gR
2 gR 3 gR 4 5 gR O gR .
70
It is easy to check that the rst two coefcients, and , are universal in the sense that for two different theories, parametrized by (g R , ) and (g R , ), the two functions have the same rst two coefcients in their expansions. This method of computing the function also lets us bypass the strange way to calculate it that we have used in Eqs. 64 and 65 , where we have rst expressed g R in terms of g 0 to calculate (g R ) as a function of g 0 and then, by inversion of the series, re-obtained a function of g R . These two steps are a priori dangerous because they both involve large logarithms. Actually, they always cancel each other. This can be seen directly for the example of Eq. 67 and the reason for this cancellation comes from Eqs. 68 and 69 , which show that no inversion of series is needed to calculate (g R ). There is no miracle here, because only the behavior at y / 1, which of course does not involve , matters. Finally, we mention that the integration of the function 3 at O(g R )analogous to a two-loop result in QFTleads to an implicit equation for g R that generalizes Eq. 66 :
physics, things are less simple. At least psychologically. It was indeed natural in the early days of quantum electrodynamics to think that this theory was fundamental, that is, not derived from a more fundamental theory. More precisely, it was believed that QED had to be mathematically internally consistent, even if in the real world new physics had to occur at higher energies. Thus, the regulator scale was introduced only as a trick to perform intermediate calculations. The limit was supposed to be the right way to eliminate this unwanted scale, which anyway seemed to have no interpretation. We shall see in the following that the community now interprets the renormalization process differently. 4 Once the theory is regularized, F can be a nontrivial function of x. The price is that different values of x now correspond to different values of the coupling constant dened as the values of F for these x). Actually, it no longer makes sense to speak of a coupling constant in itself. The only meaningful concept is the pair ( ,g R ( )) of coupling constants at a given scale. The relevant question now is, What are the physical reasons in particle/statistical physics that make the coupling constants depend on the scale while they are constants in the classical/mean eld approximation? As mentioned, for particle physics, the answer is the existence of new quantum uctuations corresponding to the possibility of creating and annihilating particles at energies higher than mc 2 . What was scale independent in the classical theory becomes scale dependent in the quantum theory because, as the available energy increases, more and more particles can be created. The pairs of virtual particles surrounding an electron are polarized by its presence and thus screen its charge. As a consequence, the charge of an electron depends on the distance or equivalently the energy at which it is probed, at least for distances smaller than the Compton wavelength. Note that the energy scale mc 2 should not be confused with the cut-off scale . mc 2 is the energy scale above which quantum uctuations start to play a signicant role while is the scale where they are cut-off. Thus, although the Compton wavelength is a short distance scale for the classical theory, it is a long distance scale for QFT, the short one being 1 . There are thus three domains of length scales in QFT: above the Compton wavelength where the theory behaves classically up to small quantum corrections coming from high energy virtual processes , between the Compton wave1 length and the cut-off scale where the relativistic and 1 where quantum uctuations play a great role, and below 12 a new, more fundamental theory has to be invoked. In statistical physics, the analog of the Compton wavelength is the correlation length which is a measure of the distance at which two microscopic constituents of the system are able to inuence each other through thermal uctuations.31 For the Ising model, for instance, the correlation length away from the critical point is the order of the lattice spacing and the corrections to the mean-eld approximation due to uctuations are small. Unlike particle physics where the masses and therefore the Compton wavelengths are xed, the correlation lengths in statistical mechanics can be tuned by varying the temperature. Near the critical temperature where the phase transition takes place, the correlation length becomes extremely large and uctuations on all length scales between 1 the microscopic scale of order , a lattice spacing, and the correlation length add up to modify the mean-eld behavior see Refs. 32, 33 and also Ref. 34 for a bibliography on this
subject . We see here a key to the relevance of renormalization: two very different scales must exist between which a nontrivial dynamics quantum or statistical in our examples can develop. This situation is a priori rather unnatural as can be seen for phase transitions, where a ne tuning of temperature must be implemented to obtain correlation lengths much larger than the microscopic scale. Most of the time, physical systems have an intrinsic scale of time, energy, length, etc. and all the other relevant scales of the problem are of the same order. All phenomena occurring at very different scales are thus almost completely suppressed. The existence of a unique relevant scale is one of the reasons why renormalization is not necessary in most physical theories. In QFT it is mandatory because the masses of the known particles are much smaller than a hypothetical cut-off scale , still to be discovered, where new physics should take place. This is a rather unnatural situation, because, contrary to phase transitions, there is no analog of a temperature that could be netuned to create a large splitting of energy, that is, mass, scales. The question of naturalness of the models we have at present in particle physics is still largely open, although there has been much effort in this direction using supersymmetry. 5 The classical theory is valid down to the Compton/ correlation length, but cannot be continued naively beyond this scale; otherwise, when mixed with the quantum formalism, it produces divergences. Actually, it is known in QFT that the elds should be considered as distributions and not as ordinary functions. The need for considering distributions comes from the nontrivial structure of the theory at very short length scale where uctuations are very important. At short distances, functions are not sufcient to describe the eld state, which is not smooth but rough, and distributions are necessary. Renormalizing the theory consists actually in building, order by order, the correct distributional continuation of the classical theory. The uctuations are then correctly taken into account and depend on the scale at which the theory is probed: this nontrivial scale dependence can only be taken into account theoretically through the dependence of the analog of the function F with x and thus of the coupling with the scale . 6 If the theory is perturbatively renormalizable, the pairs ( ,g( )) form an equivalence class of parametrizations of the theory. The change of parametrization from ( ,g( )) to ( ,g( )), called a renormalization group transformation, is then performed by a law which is self-similar, that is, such that it can be iterated several times while being form-invariant.27,30 This law is obtained by the integration of gR gR
g0 ,
73
This function has a true perturbation expansion in terms of g R unlike the perturbative relation between g R ( ) and g R ( ) which involves logarithms of / that can be large. The integration of Eq. 73 partially resums the perturbation series and is thus semi-nonperturbative even if (g R ) has been calculated perturbatively. The self-similar nature of the group law is encoded in the fact that (g R ) is independent of .5 In particle physics, the function gives the evolution of the strength of the interaction as the energy at which it is probed varies and the integration of the function resums partially the perturbation expansion. First, as the energy increases, the coupling constant can decrease and eventually
vanish. This is what happens when 0 in Eqs. 65 and 66 . In this case, the particles almost cease to interact at very high energies or equivalently when they are very close to each other. The theory is then said to be asymptotically free in the ultraviolet domain.3,5 Reciprocally, at low energies the coupling increases and perturbation theory can no longer be trusted. A possible scenario is that bound states are created at a sufciently low energy scale so that the perturbation approach has to be reconsidered in this domain to take into account these new elementary excitations. Non-Abelian gauge theories are the only known theories in four space time dimensions that are ultraviolet free, and it is widely believed that quantum chromodynamicswhich is such a theoryexplains quark connement. The other important behavior of the scale dependence of the coupling constant is obtained for 0 in which case it increases at high energies. This corresponds, for instance, to quantum electrodynamics. For this kind of theory, the dramatic increase of the coupling at high energies is supposed to be a signal that the theory ceases to be valid beyond a certain energy range and that new physics, governed by an asymptotically free theory like the standard model of electro-weak interactions , has to take place at short distances. 7 Renormalizability, or its nonperturbative equivalent, self-similarity, ensures that although the theory is initially formulated at the scale , this scale together with g 0 can be entirely eliminated for another scale better adapted to the physics we study. If the theory was solved exactly, it would make no difference which parametrization we used. However, in perturbation theory, this renormalization lets us avoid calculating small numbers as differences of very large ones. It would indeed be very unpleasant, and actually meaningless, to calculate energies of order 100 GeV, for instancethe scale of our analysisin terms of energies of order of the Planck scale 1019 GeV, the analog of the scale . In a renormalizable theory, the possibility to perturbatively eliminate the large scale has a very deep meaning: it is the signature that the physics is short distance insensitive or equivalently that there is a decoupling of the physics at different scales. The only memory of the short distance scale lies in the initial conditions of the renormalization group ow, not in the ow itself: the function does not depend on . We again emphasize that, usually, the decoupling of the physics at very different scales is trivially related to the existence of a typical scale such that the inuence of all phenomena occurring at different scales is almost completely suppressed. Here, the decoupling is much more subtle because there is no typical length in the whole domain of length scales that is very small compared with the Compton 1 . Because inwavelength and very large compared with teractions among particles correspond to nonlinearities in the theories, we could naively believe that all scales interact with each otherwhich is trueso that calculating, for instance, the low energy behavior of the theory would require the detailed calculation of all interactions occurring at higher energies. Needless to say that in a eld theory, involving innitely many degrees of freedomthe value of the eld at each pointsuch a calculation would be hopeless, apart from exactly solvable models. Fortunately, such a calculation is not necessary for physical quantities that can be calculated from renormalizable couplings only. Starting at very high energies, typically , where all coupling constants are naturally of order 1, the renormalization group ow drives almost all of them to zero, leaving only, at low energies, the renor180 Am. J. Phys., Vol. 72, No. 2, February 2004
malizable couplings. This is the interpretation of nonrenormalizable couplings. They are not terrible monsters that should be forgotten as was believed in the early days of QFT. They are simply couplings that the RG ow eliminates at low energies. If we are lucky, the renormalizable couplings become rather small after their RG evolution between and the scale at which we work, and perturbation theory is valid at this scale. We see here the phenomenon of universality: among the innitely many coupling constants that are a priori necessary to encode the dynamics of the innitely many degrees of freedom of the theory, only a few ones are nally relevant.35 All the others are washed out at large distances. This is the reason why, perturbatively, it is not possible to keep these couplings nite at large distance, and it is necessary to set them to zero.36 The simplest nontrivial example of universality is given by the law of large numbers the central limit theorem which is crucial in statistical mechanics.32 In systems where it can be applied, all the details of the underlying probability distribution of the constituents of the system are irrelevant for the cooperative phenomena which are governed by a Gaussian probability distribution.37 This drastic reduction of complexity is precisely what is necessary for physics because it lets us build effective theories in which only a few couplings are kept.12 Renormalizability in statistical eld theory is one of the nontrivial generalizations of the central limit theorem. 8 The cut-off , rst introduced as a mathematical trick to regularize integrals, has actually a deep physical meaning: it is the scale beyond which new physics occur and below which the model we study is a good effective description of the physics. In general, it involves only the renormalizable couplings and thus cannot pretend to be an exact description of the physics at all scales. However, if is very large compared with the energy scale in which we are interested, all nonrenormalizable couplings are highly suppressed and the effective model, retaining only renormalizable couplings, is valid and accurate the Wilson RG formalism is well suited to this study, see Refs. 35 and 38 . In some modelsthe asymptotically free onesit is possible to formally take the limit both perturbatively and nonperturbatively, and there is therefore no reason to invoke a more fundamental theory taking over at a nite but large . Let us emphasize here several interesting points. i For a theory corresponding to the pair ( ,g R ( )), the limit must be taken within the equivalence class of parametrizations to which ( ,g R ( )) belongs.39 A divergent nonregularized perturbation expansion consists in taking while keeping g 0 nite. From this viewpoint, the origin of the divergences is that the pair ( ,g 0 ) does not belong to any equivalence class of a sensible theory. Perturbative renormalization consists in computing g 0 as a formal powers series in g R at nite , so that ( ,g 0 ) corresponds to a mathematically consistent theory; we then take the limit . Because of universality, it is physically impossible to know from low energy data if is very large or truly innite. Although mathematically consistent, it seems unnatural to reverse the RG ow while keeping only the renormalizable couplings and thus to imagine that even at asymptotically high energies, Nature has used
ii iii
only the couplings that we are able to detect at low energies. It seems more natural that a fundamental theory does not suffer from renormalization problems. String theory is a possible candidate.40 To conclude, we see that although the renormalization procedure has not evolved much these last thirty years, our interpretation of renormalization has drastically changed:12 the renormalized theory was assumed to be fundamental, while it is now believed to be only an effective one; was interpreted as an articial parameter that was only useful in intermediate calculations, while we now believe that it corresponds to a fundamental scale where new physics occurs; nonrenormalizable couplings were thought to be forbidden, while they are now interpreted as the remnants of interaction terms in a more fundamental theory. Renormalization group is now seen as an efcient tool to build effective low energy theories when large uctuations occur between two very different scales that change the physics qualitatively and quantitatively. ACKNOWLEDGMENTS The author thanks S. Dusuel and B. Doucot for many in teresting discussions and remarks. He also thanks J. Bartlett for his help in the choice of title of this article, and L. Canet, A. Elkharrat, E. Huguet, A. Laverne, P. Lecheminant, R. Mosseri, D. Mouhanna, M. Tissier, and N. Wschebor for a careful reading of the manuscript and many suggestions. He also thanks all the students of the Ninth Vietnam School of Physics Hue, January 2003 for having worked hard on this article and for their suggestions. Laboratoire de Physique Theorique et Hautes Energies is a Laboratoire associe au CNRS: UMR 7589. APPENDIX A: TOY MODELS FOR RENORMALIZABLE AND NONRENORMALIZABLE PERTURBATION EXPANSIONS We give an example of a nonrenormalizable theory and of a theory which needs two couplings to be renormalized. Let us suppose that F 1, x dt
1
which is still logarithmically divergent for all x 0. The difference between the two examples given in Eqs. 5 and A1 is that in the last one, once the linear divergence has been subtracted, the logarithmic sub-divergence remains. Subtracting it would require us to impose a second prescription that would dene a new coupling constant. In the absence of this second coupling constant, the logarithmic divergence cannot be subtracted and the model is nonrenormalizable. Let us examine how a second coupling constant could solve the problem. Generically, this second coupling, which we call 0 , already contributes at rst order. We take as an example F x g0
0x
g2 0
dt
1
t t x
O g3 . 0
A5
Let us take as renormalization prescriptions, F x x 0

R,
A6
in addition to Eq. A2 . We obtain at rst order that g 0 2 2 g R O(g R ) and 0 R O(g R ) and at second order
2g 2 gR
dt,
1
A7
2 gR
dt . t
A8
If we substitute these expressions in Eq. A5 , we nd F x gR

Rx 2 g Rx 2
dt t t x
3 O gR .
A9
t t x
A1
which, unlike the example of Eq. 5 , is linearly divergent. To renormalize this function, we have to impose a prescription at one point, and we choose F 0 gR . A2 Note that it was not possible in the example of Eq. 5 to take 0, because this choice would have lead to a divergence of the integral at the lower bound. In Eq. A1 taking 0 is possible because the lower bound of the integral is nonvanishing and actually plays somewhat the role of a nonvanishing . We have
2g 2 gR
Obviously, this expression converges when . The two renormalization prescriptions let us subtract the linear divergence as well as the logarithmic sub-divergence. We emphasize that in the previous example we only eliminated the second divergence at order g 2 . At higher orders, there are 0 two ways a theory can behave, characterized by two different renormalizability properties. The rst one is that all divergences can be removed to all orders by renormalizing only the two couplings g 0 and 0 . A variant of this possibility is that a third couplingor a nite number of new couplings turns out to be necessary and sufcient to remove the divergences. In this case, the model is renormalizable at the price of introducing all the necessary couplings. The second possibility is that the new interaction term, which has induced the existence of the 0 term in F, generates new divergences at higher orders. In this case, new interaction terms and coupling constants are required to remove the new divergences. These new terms can themselves generate new divergences at even higher orders, which require new couplings to be removed and so on and so forth. In this case, innitely many interaction terms are necessary to remove the divergences and the model is perturbatively nonrenormalizable. APPENDIX B: DERIVATION OF EQ. 22 Let us show that it is always possible to make the choice used in Eq. 22 . Due to condition 15 , we have generally
dt,
1
A3
so that F
181
gR
2 g Rx
dt , t x
A4
s F 1, x
s F 1,
g 1 x,
when
B1
181
Am. J. Phys., Vol. 72, No. 2, February 2004
Bertrand Delamotte
where the limit g 1 is a well-dened function satisfying s g 1 ( ,x). If we rst evaluate F 1, (x) in Eq. g 1 (x, ) B1 at x 1 for instance and at 1 and subtract them, we obtain that g 1 has the following form: g 1 x, g1 x g1 , B2 namely, a combination of the same function of x and . s s s Then, by redening F 1, : F 1, F 1, g 1 we satisfy Eq. 22 . Note that the previous choice of singular part is not necessary and is only convenient.
g0 gR
2 g R log
2 3 gR
log2
3 g RG
4 O gR .
D4
Thus, substituting this expression for g 0 in g R g R (g 0 ), Eq. D2 , we obtain gR gR log

2 g R log 3 gR 2 2
log2 G . D5
log
APPENDIX C: LOGARITHMIC DIVERGENCES IN RENORMALIZABLE THEORIES WITH DIMENSIONLESS COUPLINGS We prove for renormalizable theories with dimensionless s couplings that F 1, (x) must be a logarithm. If we use Eq. 23 , dimensional analysis, and the freedom to choose the regular part of F 1, , we have
s F 1,
The group law is obeyed at this order if the relation between g R and g R is of the same form as the one between g R and g 0 , Eq. D3 . This condition requires gR gR and thus 2
2 2 g R log 3 g RG 4 O gR ,
D6
log
log
D7
f1
r x .
C1
Note that in full generality, the regular part we add to f 1 ( ) could depend on : r (x). However, because it is regular, we can choose to add only the -independent function corresponding to the limit of r : r(x) r (x). If we differentiate Eq. C1 with respect to x and then take x 1 and 1/y, we obtain f y and thus f y log y, C3 where the minus sign has been written for convenience. From C1 and C3 we conclude that f (x) r(x) f 1 (x) and that
s F 1, x
By differentiating this relation with respect to , we nd, setting x / : ing G x 2

2
and by tak-
log x x
G 1 . x
D8
If we take into account that G(1) 0, we nd by integration G x where

2
log2 x
log x,
D9
r 1 , y
C2
is arbitrary.
APPENDIX E: THE RENORMALIZATION GROUP APPLIED TO A DIFFERENTIAL EQUATION We show how the renormalization program can be implemented for the example of the differential equation 42 whose exact solution is y t f r 0 ,t t 0 r 0e
(t t 0 )
E1
f x
log
C4
In perturbation theory, we nd
2
APPENDIX D: RENORMALIZATION GROUP IMPROVED EXPANSION We show how to derive Eq. 59 . Consider the denition of F imp: F imp x,g 0 , g0 g 2 log 0 x g 3G 0 x O g4 . 0 D1 We can calculate g R and g R from their denitions where F imp is used instead of F) and from Eq. D1 : gR g0 gR g0 g 2 log 0 g 2 log 0 g 3G 0 g 3G 0 O g4 , 0 O g4 . 0 D2 D3
y t
r0 1
t t0
t t0
E2
If we invert the series g R g R (g 0 ) of Eq. D2 , we obtain

Fig. 2. The curve y(t) as a function of t. The thick lower line is the approximation of order , see Eq. E2 . The other lines represent the eld of tangent vectors to the curvethe envelopegiven by the function, Eq. E7 . Bertrand Delamotte 182
At order 0 , y(t) is constant and nite, whereas, at any higher order in , a divergence occurs for t 0 . This divergence arises of course from the fact that the expansion turns out to be in powers of (t t 0 ) and not in powers of alone the secular problem . Thus, as shown in Fig. 2, the approximation of order O( ) becomes worse and worse as t increases. A renormalization prescription consists here in imposing that for a nite : y r . E3
If we perform the calculation to order , we nd to rst order: r r0 1 t0 O

2
E4
and thus, as expected, y t r 1 t O

2
E5
The theory is perturbatively renormalizable at this order because by imposing a single renormalization prescription, it is possible to completely eliminate t 0 and r 0 . Let us dene the -function for r by r We nd r r O
2
r
r 0 ,t 0
r ,
0
E6
E7
It is very instructive to perform this calculation at higher orders because we then nd that the O( ) result of Eq. E7 is exact this result is trivially shown using the second equality of Eq. E6 . Thus, there is no O( 2 ) corrections to (r ). This result means that, in this example, there is no sub-leading terms such as n (t t 0 ) p with p n in the perturbation expansion. Clearly, the function gives the tangent to the curve y(t). Equation E7 shows that contrary to y(t), the function has a true expansion involving only one term in this example . This result is reminiscent of what we have already observed in our general discussion, see Eqs. 36 and 70 . Of course, this example is too simple because using the function leads to the same differential equation for r as the one for y(t) that we started with, Eq. 42 : the RG does not help us to solve ordinary differential equations. However, although mathematically trivial, our analysis shows that perturbation theory should not be used for large t t 0 , but that it is perfectly valid for innitesimal time steps, see Fig. 2. It also shows that the higher order terms of the perturbation expansion are completely analogous to the series of the leading logarithms we have previously encountered: they are entirely determined by the O( ) term together with self-similarity encoded in the function . Note nally that for partial differential equations PDE that describe the dynamics of innitely many degrees of freedom as in eld theory , the RG techniques do not let us reconstruct the PDE from the rst orders of perturbation theory. The functions lead to ordinary differential equations, the integration of which let us improve the perturbation computation of several quantities thanks to a partial resummation of the perturbation expansion.27,30
Electronic mail: delamotte@lpthe.jussieu.fr H. Bethe, The electromagnetic shift of energy levels, Phys. Rev. 72, 339341 1947 . 2 S. Weinberg, The Quantum Theory of Fields Cambridge U.P., Cambridge, 1995 . 3 J. Collins, Renormalization Cambridge U.P., Cambridge, 1984 . 4 L. Ryder, Quantum Field Theory Cambridge U.P., Cambridge, 1985 . 5 M. Le Bellac, Quantum and Statistical Field Theory Oxford U.P., Oxford, 1992 . 6 J. Binney, N. Dowrick, A. Fisher, and M. Newman, The Theory of Critical Phenomena: An Introduction to the Renormalization Group Oxford U.P., Oxford, 1992 . 7 N. Goldenfeld, Lectures on Phase Transitions and the Renormalization Group AddisonWesley, Reading, MA, 1992 . 8 Conceptual Foundations of Quantum Field Theory, edited by T. Cao Cambridge U.P., Cambridge, 1999 . 9 K. G. Wilson and J. Kogut, The renormalization group and the -expansion, Phys. Rep. C 12, 75199 1974 . 10 Some phase transitions are triggered by quantum uctuations. This subtlety plays no role in what follows. 11 The short distance physics in statistical systems is given by Hamiltonians describing, for instance, interactions among magnetic ions or molecules of a uid. 12 G. Lepage, What is renormalization?, TASI89 Summer School, Boulder, ASI, 1989, p. 483. 13 P. Kraus and D. Grifths, Renormalization of a model quantum eld theory, Am. J. Phys. 60, 10131023 1992 . 14 I. Mitra, A. DasGupta, and B. Dutta-Roy, Regularization and renormalization in scattering from Dirac delta-potentials, Am. J. Phys. 66, 1101 1109 1998 . 15 P. Gosdzinsky and R. Tarrach, Learning quantum eld theory from elementary quantum mechanics, Am. J. Phys. 59, 7074 1991 . 16 L. Mead and J. Godines, An analytical example of renormalization in two-dimensional quantum mechanics, Am. J. Phys. 59, 935937 1991 . 17 K. Adhikari and A. Ghosh, Renormalization in non-relativistic quantum mechanics, J. Phys. A 30, 6553 6564 1997 . 18 Actually, the analog of the function in Eq. 1 would be a correlation function of four density or spin elds taken at four different points. These functions are not easily measurable and thus g 0 does not have in general an intuitive meaning in this case. Because this subtlety plays no role in our discussion, we ignore this difculty in the following. 19 Actually for QED it is a four-dimensional integral over four-momenta and the integrand is a product of propagators. 20 R. Delbourgo, How to deal with innite integrals in quantum eld theory, Rep. Prog. Phys. 39, 345399 1976 . 21 S. Nyeo, Regularization methods for delta-function potential in twodimensional quantum mechanics, Am. J. Phys. 68, 571575 2000 . 22 It is nontrivial to prove in full generality that the results obtained after renormalization are independent of the regularization scheme. However, it is easy to grasp the idea behind it. Because renormalization consists in eliminating parameters like g 0 and in replacing them by measurable couplings like g R , the renormalized quantities like F(x) are nally expressed only in terms of physical quantities that are independent of the regularization scheme Ref. 15 . 23 M. Hans, An electrostatic example to illustrate dimensional regularization and renormalization group technique, Am. J. Phys. 51, 694 698 1983 . 24 Note that the renormalized series in g R can themselves be nonconvergent. Most of the time they are at best asymptotic. In some cases they can be resummed using Borel transform and Pade approximants. 25 In QFT, it is in general also necessary to change the normalization of the analog of the function Fthe Green functionsby a factor that diverges in the limit . This procedure is known as eld renormalization. 26 Let us emphasize that there is a subtlety if dimensional regularization is used. Actually, this regularization also introduces a dimensional parameter , which is not directly a regulator as is the cut-off in the integral of Eq. 5 . The analog of in this regularization is given by exp(1/ ), where 4 D and D is the spatial dimension. It often is convenient to take . We mention that dimensional regularization kills all nonlogarithmic divergences Ref. 14 . 27 D. Shirkov and V. Kovalev, Bogoliubov renormalization group and symmetry of solutions in mathematical physics, Phys. Rep. 352, 219249 2001 . 28 The elements of the group are the functions: g t f (,t) for t R. They
1
Bertrand Delamotte
183
transform an initial condition r 0 into the solution at a later time interval t of the differential equation we consider see Eq. 43 in our example : g t (r 0 ) f (r 0 ,t). The composition law is thus g t .g t f ( f (,t),t ). It obeys trivially the identity: g t .g t g t t which is nothing but Eq. 44 . The identity is g t 0 and the inverse is g t . The law is associative as it should be for a group. 29 If we had not omitted in Eq. 52 the nite parts, we would have found F(x ) g 0 ag 2 bg 3 . Thus g 0 is in general not associated ex0 0 actly with the scale , but with up to a factor of order unity. 30 V. Kovalev and D. Shirkov, Functional self-similarity and renormalization group symmetry in mathematical physics, Theor. Math. Phys. 121, 13151332 1999 . 31 It is quite similar to the Compton wavelength of the pion which is the typical range of the nuclear force between hadrons like protons and nucleons. 32 E. Raposo, S. de Oliveira, A. Nemirovsky, and M. Coutinho-Filho, Random walks: A pedestrian approach to polymers, critical phenomena and eld theory, Am. J. Phys. 59, 633 645 1991 . 33 H. Stanley, Scaling, universality, and renormalization: Three pillars of modern critical phenomena, Rev. Mod. Phys. 71, S358 S366 1999 .
34
J. Tobochnik, Resource Letter CPPPT-1: Critical point phenomena and phase transitions, Am. J. Phys. 69, 255263 2001 . 35 C. Bagnuls and C. Bervillier, Exact renormalization group equations: An introductory review, Phys. Rep. 348, 91157 2001 . 36 More precisely, working with nonrenormalizable couplings would 1 to unnatural require us to ne-tune innitely many of them at scale values. Most of the time, such a nely-tuned model is no longer predictive. 37 G. Jona-Lasinio, Renormalization group and probability theory, cond-mat/0009219. 38 N. Tetradis and C. Wetterich, Critical exponents from the effective average action, Nucl. Phys. B FS 422, 541592 1994 . 39 Nonperturbatively, the existence of the limit is more subtle than perturbatively. The renormalization group ow must be controlled in this limit and this is achieved if nonperturbatively g 0 has a nite limit, that is, if there exists an ultraviolet xed point of the RG ow. The case g 0 0 when , corresponds to asymptotically free theories, that is, in four spacetime dimensions, to non-Abelian gauge theories. 40 C. Schmidhuber, On water, steam, and string theory, Am. J. Phys. 65, 10421050 1997 .
184
Am. J. Phys., Vol. 72, No. 2, February 2004
Bertrand Delamotte
184

(AJP) A Hint of Re Normalization

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(AJP) A Hint of Re Normalization

Uploaded by

Copyright:

Available Formats

A hint of renormalization

Thus the use of Eq. 3 leads to

where we have used at this order, we obtain

which diverges when nd

. In our example, Eq. 5 , we dt

If we substitute Eq. 12 into Eq. 11 , we obtain F to this order: F x

If we substitute Eq. 18 in Eq. 17 , we obtain F x

x-independent. In our example of Eqs. 2 and 24 we nd

In our example, Eq. 5 , we can choose f1 log ,

25 Equation 25 can be rewritten as

s Thus, we obtain for the new function F 2, (x):

By substituting this expression in Eqs. 36 and 39 , we 4 obtain at O(g R ): F x gR

the solution of which is f r 0 ,t t 0 The group law can be written as f r 0 ,t t 0 f f r0 ,

we obtain gR g0 gR g0 g 2 log 0 g 2 log 0 O g3 , 0 O g3 . 0 55 56

By imposing the group law, Eq. 44 , to order a functional equation for G: G t t0 G t0 G t t0 t

By eliminating g 0 between these two equations, we nd gR gR

If we differentiate Eq. 48 with respect to t 0 and take t 0 , we obtain, setting x t , G x x G 0 . x2 2 49

Because G(0) 0, Eq. 49 implies that G x ax, 50

is arbitrary. Thus, g0 g 2 log 0 g 3 log 0

Now, if we integrate Eq. 63 together with Eq. 65 , we obtain gR 1 gR g R log . 66

If f is a double series in g and in log( / f gR ,

Let us take as renormalization prescriptions, F x x 0

If we substitute these expressions in Eq. A5 , we nd F x gR

Am. J. Phys., Vol. 72, No. 2, February 2004

Thus, substituting this expression for g 0 in g R g R (g 0 ), Eq. D2 , we obtain gR gR log

By differentiating this relation with respect to , we nd, setting x / : ing G x 2

If we take into account that G(1) 0, we nd by integration G x where

If we invert the series g R g R (g 0 ) of Eq. D2 , we obtain

If we perform the calculation to order , we nd to rst order: r r0 1 t0 O

and thus, as expected, y t r 1 t O

Am. J. Phys., Vol. 72, No. 2, February 2004

You might also like